Keep every base's slots under multiple inheritance

[ESultanik]

Closes #173.

Builds on @Liyu0310-Code's work in #175, which was merged into this branch so their commit is carried over intact, and finishes it.

The bug

DataClassNode.__init_subclass__ accumulated slots by reading the inherited _SLOTS / _SLOT_ANNOTATIONS attributes, which resolve through the MRO to only the first inheritance chain. Under diamond inheritance the slots of every base after the first were silently dropped, so items(), __iter__, to_obj() and the hash all ignored them, and constructing the class failed with a misleading TypeError: object.__init__() takes exactly one argument.

The fix collects annotations from every data-class ancestor and derives _SLOTS from the merged mapping. The reproducer in #173 now gives:

D._SLOTS = ('a', 'c', 'b', 'd')
to_obj   = {'a': 1, 'c': 3, 'b': 2, 'd': 4}

What this branch corrects

df537d6 fixes two things in the merged change:

The ancestors were iterated in the wrong order. The slot merge was written against an earlier __init_subclass__, where ancestors was in plain __mro__ order and had to be reversed to merge base-first. #166 has since rebuilt that list from reversed(cls.__mro__), so it already arrives base-first, and reversing it again inverted the merge order — a diamond class D(B, C) came out in B-then-C order. None of the edited lines overlapped with #166's, so the change applied cleanly while its meaning changed underneath it, and test_multiple_inheritance failed on the merge commit. The test itself was right; only the implementation disagreed with it.

new_slots had become dead. Once _SLOTS is derived from the merged annotations, the list is written but never read. Ruff does not report it: F841 wants a local that is never referenced, and new_slots.append(...) counts as a reference.

Validation

• test_multiple_inheritance fails on d7c0f6a (the merge of Fix DataClassNode slot accumulation under multiple inheritance #175) and passes here. Reinstating the extra reversed() reproduces the failure on this branch, so the test is pinned to the defect rather than passing incidentally.
• uv run --frozen --extra dev pytest — 175 passed, 19 subtests
• uv run --frozen --extra dev ruff check graphtage test docs bindist — clean
• uv run --frozen --extra dev make -C docs html SPHINXOPTS="-W --keep-going" — build succeeded

The full suite passing matters here beyond the new test: every in-tree DataClassNode is constructed positionally by its builder, so a change to _SLOTS ordering would surface as widespread failures rather than quietly.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GypKU5KdLfs2Cf8kS2TzJa